core-3/kuno-dogwalker-7b
core-3/kuno-dogwalker-7b is a 7 billion parameter language model created by core-3, formed by merging SanjiWatsuki/Kunoichi-DPO-v2-7B and mlabonne/AlphaMonarch-7B using LazyMergekit. This model demonstrates competitive benchmark performance across reasoning, common sense, and language understanding tasks, particularly excelling in GSM8K math problems. It is suitable for general-purpose text generation where strong mathematical reasoning is beneficial.
Loading preview...
Overview
core-3/kuno-dogwalker-7b is a 7 billion parameter merged language model developed by core-3. It was created using LazyMergekit by combining two base models: SanjiWatsuki/Kunoichi-DPO-v2-7B and mlabonne/AlphaMonarch-7B. The merge utilized a slerp method with specific t parameters for self-attention and MLP layers, configured in bfloat16.
Key Capabilities & Performance
kuno-dogwalker-7b exhibits strong performance across various benchmarks, achieving an average score of 74.94. Notably, it scores 71.11 on GSM8K, indicating solid mathematical reasoning capabilities. Its performance is competitive with, and in some cases surpasses, other 7B models like core-3/kuno-royale-v2-7b and SanjiWatsuki/Kunoichi-DPO-v2-7B across metrics such as ARC, HellaSwag, MMLU, and Winogrande.
Differentiators & Use Cases
While its overall metrics are decent, the model's developers note that its "writing feels off" compared to core-3/kuno-royale-v2-7b, suggesting potential nuances in its generative style. Despite this, its strong GSM8K score makes it a good candidate for applications requiring numerical or logical problem-solving. Developers can integrate it using the provided Hugging Face transformers pipeline for text generation tasks.